Sim2Real When Data Is Scarce: Image Transformation for Industrial Applications

Synthetic data for training deep neural networks is increasingly used in computer vision. Several strategies, such as domain randomization or domain adaptation (sim2real), exist to bridge the domain gap between synthetic training data and the real application. We compare different image transformation models, such as generative adversarial networks (GANs), to adapt the synthetic images to some real-world examples. Our focus is on the transfer capability and the consistency of the annotations. We investigate the influence of different augmentation strategies on the transformation capability. Our study is exemplified by an industrial quality assurance use case. We show that especially a strong translational augmentation is a key for a successful sim2real transfer using GANs. Trained on a transformed dataset, our object detectors achieve an almost equivalent performance compared to real-world data. Based on the industrial use case, we even prove the superiority of synthetic image data for quality assurance.